fabi/Xenia-Canary
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #209 from DrChat/muladdsub_cleanup

Browse Source 
Rewrite FMA mul sub/add to avoid register shuffling
This commit is contained in:
Ben Vanik
2015-05-16 13:25:28 -07:00
parent 504b855c20 5dbec09818
commit 3a7d1f21e8
1 changed files with 56 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files

110
src/xenia/cpu/backend/x64/x64_sequences.cc
View File

@@ -3742,21 +3742,24 @@ EMITTER_OPCODE_TABLE(
// TODO(benvanik): use other forms (132/213/etc) to avoid register shuffling.
// dest could be src2 or src3 - need to ensure it's not before overwriting dest
// perhaps use other 132/213/etc
// Forms:
// - 132 -> $1 = $1 * $3 + $2
// - 213 -> $1 = $2 * $1 + $3
// - 231 -> $1 = $2 * $3 + $1
EMITTER(MUL_ADD_F32, MATCH(I<OPCODE_MUL_ADD, F32<>, F32<>, F32<>, F32<>>)) {
static void Emit(X64Emitter& e, const EmitArgType& i) {
// FMA extension
if (e.IsFeatureEnabled(kX64EmitFMA)) {
if (i.dest == i.src1) {
e.vfmadd213ss(i.dest, i.src2, i.src3);
} else if (i.dest == i.src2) {
e.vfmadd213ss(i.dest, i.src1, i.src3);
} else if (i.dest == i.src3) {
e.vfmadd231ss(i.dest, i.src1, i.src2);
} else {
if (i.dest != i.src2 && i.dest != i.src3) {
e.vmovss(i.dest, i.src1);
e.vfmadd213ss(i.dest, i.src2, i.src3);
} else {
e.vmovss(e.xmm0, i.src1);
e.vfmadd213ss(e.xmm0, i.src2, i.src3);
e.vmovss(i.dest, e.xmm0);
}
// Dest not equal to anything
e.vmovss(i.dest, i.src1);
e.vfmadd213ss(i.dest, i.src2, i.src3);
}
} else {
// If i.dest == i.src3, back up i.src3 so we don't overwrite it.
@@ -3767,7 +3770,7 @@ EMITTER(MUL_ADD_F32, MATCH(I<OPCODE_MUL_ADD, F32<>, F32<>, F32<>, F32<>>)) {
}
e.vmulss(i.dest, i.src1, i.src2); // $0 = $1 * $2
e.addss(i.dest, src3); // $0 = $0 + $1
e.vaddss(i.dest, i.dest, src3); // $0 = $1 + $2
}
}
};
@@ -3777,15 +3780,14 @@ EMITTER(MUL_ADD_F64, MATCH(I<OPCODE_MUL_ADD, F64<>, F64<>, F64<>, F64<>>)) {
if (e.IsFeatureEnabled(kX64EmitFMA)) {
if (i.dest == i.src1) {
e.vfmadd213sd(i.dest, i.src2, i.src3);
} else if (i.dest == i.src2) {
e.vfmadd213sd(i.dest, i.src1, i.src3);
} else if (i.dest == i.src3) {
e.vfmadd231sd(i.dest, i.src1, i.src2);
} else {
if (i.dest != i.src2 && i.dest != i.src3) {
e.vmovsd(i.dest, i.src1);
e.vfmadd213sd(i.dest, i.src2, i.src3);
} else {
e.vmovsd(e.xmm0, i.src1);
e.vfmadd213sd(e.xmm0, i.src2, i.src3);
e.vmovsd(i.dest, e.xmm0);
}
// Dest not equal to anything
e.vmovsd(i.dest, i.src1);
e.vfmadd213sd(i.dest, i.src2, i.src3);
}
} else {
// If i.dest == i.src3, back up i.src3 so we don't overwrite it.
@@ -3796,7 +3798,7 @@ EMITTER(MUL_ADD_F64, MATCH(I<OPCODE_MUL_ADD, F64<>, F64<>, F64<>, F64<>>)) {
}
e.vmulsd(i.dest, i.src1, i.src2); // $0 = $1 * $2
e.addsd(i.dest, src3); // $0 = $0 + $1
e.vaddsd(i.dest, i.dest, src3); // $0 = $1 + $2
}
}
};
@@ -3806,15 +3808,14 @@ EMITTER(MUL_ADD_V128, MATCH(I<OPCODE_MUL_ADD, V128<>, V128<>, V128<>, V128<>>))
if (e.IsFeatureEnabled(kX64EmitFMA)) {
if (i.dest == i.src1) {
e.vfmadd213ps(i.dest, i.src2, i.src3);
} else if (i.dest == i.src2) {
e.vfmadd213ps(i.dest, i.src1, i.src3);
} else if (i.dest == i.src3) {
e.vfmadd231ps(i.dest, i.src1, i.src2);
} else {
if (i.dest != i.src2 && i.dest != i.src3) {
e.vmovdqa(i.dest, i.src1);
e.vfmadd213ps(i.dest, i.src2, i.src3);
} else {
e.vmovdqa(e.xmm0, i.src1);
e.vfmadd213ps(e.xmm0, i.src2, i.src3);
e.vmovdqa(i.dest, e.xmm0);
}
// Dest not equal to anything
e.vmovdqa(i.dest, i.src1);
e.vfmadd213ps(i.dest, i.src2, i.src3);
}
} else {
// If i.dest == i.src3, back up i.src3 so we don't overwrite it.
@@ -3825,7 +3826,7 @@ EMITTER(MUL_ADD_V128, MATCH(I<OPCODE_MUL_ADD, V128<>, V128<>, V128<>, V128<>>))
}
e.vmulps(i.dest, i.src1, i.src2); // $0 = $1 * $2
e.addps(i.dest, src3); // $0 = $0 + $1
e.vaddps(i.dest, i.dest, src3); // $0 = $1 + $2
}
}
};
@@ -3844,21 +3845,24 @@ EMITTER_OPCODE_TABLE(
// TODO(benvanik): use other forms (132/213/etc) to avoid register shuffling.
// dest could be src2 or src3 - need to ensure it's not before overwriting dest
// perhaps use other 132/213/etc
// Forms:
// - 132 -> $1 = $1 * $3 - $2
// - 213 -> $1 = $2 * $1 - $3
// - 231 -> $1 = $2 * $3 - $1
EMITTER(MUL_SUB_F32, MATCH(I<OPCODE_MUL_SUB, F32<>, F32<>, F32<>, F32<>>)) {
static void Emit(X64Emitter& e, const EmitArgType& i) {
// FMA extension
if (e.IsFeatureEnabled(kX64EmitFMA)) {
if (i.dest == i.src1) {
e.vfmsub213ss(i.dest, i.src2, i.src3);
} else if (i.dest == i.src2) {
e.vfmsub213ss(i.dest, i.src1, i.src3);
} else if (i.dest == i.src3) {
e.vfmsub231ss(i.dest, i.src1, i.src2);
} else {
if (i.dest != i.src2 && i.dest != i.src3) {
e.vmovss(i.dest, i.src1);
e.vfmsub213ss(i.dest, i.src2, i.src3);
} else {
e.vmovss(e.xmm0, i.src1);
e.vfmsub213ss(e.xmm0, i.src2, i.src3);
e.vmovss(i.dest, e.xmm0);
}
// Dest not equal to anything
e.vmovss(i.dest, i.src1);
e.vfmsub213ss(i.dest, i.src2, i.src3);
}
} else {
// If i.dest == i.src3, back up i.src3 so we don't overwrite it.
@@ -3869,7 +3873,7 @@ EMITTER(MUL_SUB_F32, MATCH(I<OPCODE_MUL_SUB, F32<>, F32<>, F32<>, F32<>>)) {
}
e.vmulss(i.dest, i.src1, i.src2); // $0 = $1 * $2
e.subss(i.dest, src3); // $0 = $0 - $1
e.vsubss(i.dest, i.dest, src3); // $0 = $1 - $2
}
}
};
@@ -3879,15 +3883,14 @@ EMITTER(MUL_SUB_F64, MATCH(I<OPCODE_MUL_SUB, F64<>, F64<>, F64<>, F64<>>)) {
if (e.IsFeatureEnabled(kX64EmitFMA)) {
if (i.dest == i.src1) {
e.vfmsub213sd(i.dest, i.src2, i.src3);
} else if (i.dest == i.src2) {
e.vfmsub213sd(i.dest, i.src1, i.src3);
} else if (i.dest == i.src3) {
e.vfmsub231sd(i.dest, i.src1, i.src2);
} else {
if (i.dest != i.src2 && i.dest != i.src3) {
e.vmovsd(i.dest, i.src1);
e.vfmsub213sd(i.dest, i.src2, i.src3);
} else {
e.vmovsd(e.xmm0, i.src1);
e.vfmsub213sd(e.xmm0, i.src2, i.src3);
e.vmovsd(i.dest, e.xmm0);
}
// Dest not equal to anything
e.vmovsd(i.dest, i.src1);
e.vfmsub213sd(i.dest, i.src2, i.src3);
}
} else {
// If i.dest == i.src3, back up i.src3 so we don't overwrite it.
@@ -3898,7 +3901,7 @@ EMITTER(MUL_SUB_F64, MATCH(I<OPCODE_MUL_SUB, F64<>, F64<>, F64<>, F64<>>)) {
}
e.vmulsd(i.dest, i.src1, i.src2); // $0 = $1 * $2
e.subsd(i.dest, src3); // $0 = $0 - $1
e.vsubsd(i.dest, i.dest, src3); // $0 = $1 - $2
}
}
};
@@ -3908,15 +3911,14 @@ EMITTER(MUL_SUB_V128, MATCH(I<OPCODE_MUL_SUB, V128<>, V128<>, V128<>, V128<>>))
if (e.IsFeatureEnabled(kX64EmitFMA)) {
if (i.dest == i.src1) {
e.vfmsub213ps(i.dest, i.src2, i.src3);
} else if (i.dest == i.src2) {
e.vfmsub213ps(i.dest, i.src1, i.src3);
} else if (i.dest == i.src3) {
e.vfmsub231ps(i.dest, i.src1, i.src2);
} else {
if (i.dest != i.src2 && i.dest != i.src3) {
e.vmovdqa(i.dest, i.src1);
e.vfmsub213ps(i.dest, i.src2, i.src3);
} else {
e.vmovdqa(e.xmm0, i.src1);
e.vfmsub213ps(e.xmm0, i.src2, i.src3);
e.vmovdqa(i.dest, e.xmm0);
}
// Dest not equal to anything
e.vmovdqa(i.dest, i.src1);
e.vfmsub213ps(i.dest, i.src2, i.src3);
}
} else {
// If i.dest == i.src3, back up i.src3 so we don't overwrite it.
@@ -3927,7 +3929,7 @@ EMITTER(MUL_SUB_V128, MATCH(I<OPCODE_MUL_SUB, V128<>, V128<>, V128<>, V128<>>))
}
e.vmulps(i.dest, i.src1, i.src2); // $0 = $1 * $2
e.subps(i.dest, i.src3); // $0 = $0 - $1
e.vsubps(i.dest, i.dest, src3); // $0 = $1 - $2
}
}
};